Species distribution models (SDMs) link species observations to environmental variables and are crucial tools for biodiversity conservation. Methods for collecting environmental variables generally consist of field-based methods or basic summaries of remotely sensed data. Basic summaries of remotely sensed data (e.g., mean and standard deviation of image bands) are not sufficient for capturing elaborate habitat characteristics and are rudimentary when compared to state-of-the-art computer vision techniques. Summarizing habitats with basic statistics is an oversimplification of landscapes and prevents models from determining how different spatial configurations of habitats impact species.
Given the advancements in computer vision due to deep learning, features extracted by deep networks have the potential to characterize habitats better than methods currently used in SDMs. To develop habitat summaries that better capture complex spatial attributes, we summarize remotely sensed data with deep neural networks. To do this, we train deep neural networks to predict land cover from remotely sensed NAIP imagery. Habitat summaries are then extracted from images with our deep networks and are subsequently fed into SDMs as habitat covariates. To test the effectiveness of our “deep” habitat summaries, we compare SDMs built with standard habitat summaries to those built with our novel “deep” habitat summaries.
Results/Conclusions
We modeled occurrences for five resident species in Oregon for the year 2011. We compared our "deep" habitat summaries to two baseline covariate sets, eBird and mean & standard deviation. The eBird covariates consist of land cover configuration summaries derived with FRAGSTATS metrics. The mean & standard deviation covariates are the simplest covariate set and summarize satellite images by the mean and standard deviation of each image band.
Surprisingly, there is little difference in model performance across the habitat-based covariate sets. There are several hypotheses for our unexpected results. First, it is possible that simple habitat summaries are sufficient indicators of species occurrences. Second, it is also possible that the generalist species we analyzed are less sensitive to habitat configuration, in which case simple habitat summaries are sufficient for predicting occurrences.
While our preliminary results do not show an advantage of habitat summaries extracted with deep networks, there is still reason to believe “deep” habitat summaries have the potential to outperform simpler habitat summaries in SDMs. We expect that our ongoing work to understand and extend these results (e.g., modeling species more sensitive to habitat type and configuration) will clarify best practices for extracting habitat summaries from remotely sensed images.